@vellumai/assistant 0.10.12 → 0.11.0-staging.2

package/src/__tests__/config-sounds-sync.test.ts

@@ -1,7 +1,7 @@
 import { describe, expect, test } from "bun:test";
 
+import type { AssistantEventEnvelope } from "../api/index.js";
 import { SYNC_TAGS } from "../daemon/message-types/sync.js";
-import type { AssistantEvent } from "../runtime/assistant-event.js";
 import { assistantEventHub } from "../runtime/assistant-event-hub.js";
 import {
   publishConfigChanged,
@@ -22,7 +22,7 @@ async function waitFor(predicate: () => boolean): Promise<void> {
 
 describe("config and sounds sync events", () => {
   test("config changes emit legacy config event and sync tag", async () => {
-    const received: AssistantEvent[] = [];
+    const received: AssistantEventEnvelope[] = [];
     const subscription = assistantEventHub.subscribe({
       type: "process",
       callback: (event) => {
@@ -48,7 +48,7 @@ describe("config and sounds sync events", () => {
   });
 
   test("sounds config changes emit legacy sounds event and sync tag", async () => {
-    const received: AssistantEvent[] = [];
+    const received: AssistantEventEnvelope[] = [];
     const subscription = assistantEventHub.subscribe({
       type: "process",
       callback: (event) => {
@@ -74,7 +74,7 @@ describe("config and sounds sync events", () => {
   });
 
   test("schedule changes emit a sync tag", async () => {
-    const received: AssistantEvent[] = [];
+    const received: AssistantEventEnvelope[] = [];
     const subscription = assistantEventHub.subscribe({
       type: "process",
       callback: (event) => {

package/src/__tests__/confirmation-request-guardian-bridge.test.ts

@@ -143,6 +143,11 @@ describe("bridgeConfirmationRequestToGuardian", () => {
     expect(emittedSignals[0].sourceEventName).toBe("guardian.question");
     expect(emittedSignals[0].sourceChannel).toBe("telegram");
     expect(emittedSignals[0].sourceContextId).toBe("conv-1");
+    // The in-app card is pinned to the conversation the confirmation was
+    // emitted in — never left to LLM conversation routing.
+    expect(emittedSignals[0].conversationAffinityHint).toEqual({
+      vellum: "conv-1",
+    });
 
     const payload = emittedSignals[0].contextPayload as Record<string, unknown>;
     expect(payload.requestId).toBe(guardianRequest.id);

package/src/__tests__/context-search-memory-source.test.ts

@@ -13,7 +13,7 @@ const embedModule = import.meta.resolve("../persistence/embeddings/embed.js");
 const embeddingBackendModule = import.meta
   .resolve("../persistence/embeddings/embedding-backend.js");
 const graphSearchModule = import.meta
-  .resolve("../plugins/defaults/memory/graph/graph-search.js");
+  .resolve("../plugins/defaults/memory/v1/graph/graph-search.js");
 const graphStoreModule = import.meta
   .resolve("../plugins/defaults/memory/graph/store.js");
 const memoryV2SourceModule = import.meta

package/src/__tests__/context-search-memory-v2-source.test.ts

@@ -26,8 +26,8 @@ import { afterEach, beforeEach, describe, expect, mock, test } from "bun:test";
 
 import type { AssistantConfig } from "../config/schema.js";
 import type { RecallSearchContext } from "../plugins/defaults/memory/context-search/types.js";
-import type { EdgeIndex } from "../plugins/defaults/memory/v3/substrate/edge-index.js";
-import type { ConceptPage } from "../plugins/defaults/memory/v3/substrate/types.js";
+import type { EdgeIndex } from "../plugins/defaults/memory/substrate/edge-index.js";
+import type { ConceptPage } from "../plugins/defaults/memory/substrate/types.js";
 let denseEmbedReturn: number[] = [0.1, 0.2, 0.3];
 mock.module("../persistence/embeddings/embedding-backend.js", () => ({
   isEmbeddingDimensionAvailable: async () => true,
@@ -48,14 +48,14 @@ let qdrantHits: QdrantHit[] = [];
 let qdrantThrows: Error | null = null;
 const qdrantCalls: Array<{ limit: number }> = [];
 
-mock.module("../plugins/defaults/memory/v3/substrate/qdrant.js", () => ({
+mock.module("../plugins/defaults/memory/substrate/qdrant.js", () => ({
   hybridQueryConceptPages: async (
     _dense: number[],
     _sparse: { indices: number[]; values: number[] },
     limit: number,
   ): Promise<QdrantHit[]> => {
     qdrantCalls.push({ limit });
-    if (qdrantThrows) throw qdrantThrows;
+    if (qdrantThrows) {throw qdrantThrows;}
     return qdrantHits;
   },
 }));
@@ -65,7 +65,7 @@ let edgeIndex: EdgeIndex = {
   incoming: new Map<string, Set<string>>(),
 };
 
-mock.module("../plugins/defaults/memory/v3/substrate/edge-index.js", () => ({
+mock.module("../plugins/defaults/memory/substrate/edge-index.js", () => ({
   getEdgeIndex: async (): Promise<EdgeIndex> => edgeIndex,
   invalidateEdgeIndex: () => {},
   getReachable: () => new Set<string>(),
@@ -75,7 +75,7 @@ mock.module("../plugins/defaults/memory/v3/substrate/edge-index.js", () => ({
 
 const pageStore = new Map<string, ConceptPage>();
 
-mock.module("../plugins/defaults/memory/v3/substrate/page-store.js", () => ({
+mock.module("../plugins/defaults/memory/substrate/page-store.js", () => ({
   getConceptsDir: (workspaceDir: string): string =>
     join(workspaceDir, "memory", "concepts"),
   listPages: async (): Promise<string[]> => [...pageStore.keys()],

package/src/__tests__/context-token-estimator.test.ts

@@ -7,7 +7,8 @@ import {
   estimatePromptTokens,
   estimateTextTokens,
 } from "../context/token-estimator.js";
-import type { Message } from "../providers/types.js";
+import { estimateOpenAICompatAudioTokens } from "../providers/openai/input-audio.js";
+import type { ContentBlock, Message } from "../providers/types.js";
 
 /** Build a minimal valid PNG header with the given dimensions, returned as base64. */
 function makePngBase64(width: number, height: number): string {
@@ -824,3 +825,55 @@ describe("tool_result estimation mirrors Anthropic wire format", () => {
     expect(inflated).toBe(wireShape);
   });
 });
+
+describe("audio-capable OpenAI-compatible models bill inline audio by duration", () => {
+  // 120 KB decoded → ~7.5 s at the compressed-audio byte-rate assumption.
+  const audioBase64 = "A".repeat(160_000);
+  const audioBytes = Math.floor((160_000 * 3) / 4);
+
+  const audioBlock = (mediaType: string): ContentBlock => ({
+    type: "file",
+    source: {
+      type: "base64",
+      media_type: mediaType,
+      data: audioBase64,
+      filename: "clip.wav",
+    },
+  });
+
+  test("catalog `supportsAudioInput` model adds the duration-based cost", () => {
+    const withModel = estimateContentBlockTokens(audioBlock("audio/wav"), {
+      providerName: "baseten",
+      model: "thinkingmachines/inkling",
+    });
+    const withoutModel = estimateContentBlockTokens(audioBlock("audio/wav"), {
+      providerName: "baseten",
+    });
+    expect(withModel - withoutModel).toBe(
+      estimateOpenAICompatAudioTokens(audioBytes),
+    );
+    expect(withModel - withoutModel).toBeGreaterThan(0);
+  });
+
+  test("non-flagged model and non-audio mime add nothing", () => {
+    const base = estimateContentBlockTokens(audioBlock("audio/wav"), {
+      providerName: "baseten",
+    });
+    expect(
+      estimateContentBlockTokens(audioBlock("audio/wav"), {
+        providerName: "baseten",
+        model: "some-other-model",
+      }),
+    ).toBe(base);
+    expect(
+      estimateContentBlockTokens(audioBlock("audio/x-m4a"), {
+        providerName: "baseten",
+        model: "thinkingmachines/inkling",
+      }),
+    ).toBe(
+      estimateContentBlockTokens(audioBlock("audio/x-m4a"), {
+        providerName: "baseten",
+      }),
+    );
+  });
+});

package/src/__tests__/conversation-agent-loop-disk-pressure.test.ts

@@ -2,9 +2,9 @@ import { afterAll, beforeEach, describe, expect, mock, test } from "bun:test";
 
 import { CompactionCircuit } from "../agent/compaction-circuit.js";
 import type { AgentLoop } from "../agent/loop.js";
+import type { AssistantEvent } from "../api/index.js";
 import type { Conversation } from "../daemon/conversation.js";
 import type { DiskPressureStatus } from "../daemon/disk-pressure-guard.js";
-import type { ServerMessage } from "../daemon/message-protocol.js";
 
 type Context = Conversation;
 
@@ -135,7 +135,16 @@ function makeCtx(overrides: Partial<Context> = {}): Conversation {
     getQueueDepth: () => 0,
     hasQueuedMessages: () => false,
     canHandoffAtCheckpoint: () => false,
-    drainQueue: async () => {},
+    drainQueue: async (_reason?: string) => {},
+    // Forwards to drainQueue so tests that spy the drain observe the agent
+    // loop's post-turn kick through the guarded entry point.
+    kickDrainQueue(
+      this: { drainQueue: (reason?: string) => unknown },
+      reason: string = "loop_complete",
+      _origin?: string,
+    ) {
+      return this.drainQueue(reason);
+    },
     getTurnInterfaceContext: () => null,
     getTurnChannelContext: () => null,
 
@@ -157,7 +166,7 @@ describe("runAgentLoopImpl disk pressure gate", () => {
   });
 
   test("blocks background turns inside the cleanup-safe finally path", async () => {
-    const events: ServerMessage[] = [];
+    const events: AssistantEvent[] = [];
     const activityStates: unknown[][] = [];
     const drainQueue = mock(async (_reason: unknown) => {});
     const ctx = makeCtx({

package/src/__tests__/conversation-agent-loop-fatal-cleanup.test.ts

@@ -17,9 +17,9 @@ import { afterAll, beforeEach, describe, expect, mock, test } from "bun:test";
 
 import { CompactionCircuit } from "../agent/compaction-circuit.js";
 import type { AgentLoop } from "../agent/loop.js";
+import type { AssistantEvent } from "../api/index.js";
 import type { Conversation } from "../daemon/conversation.js";
 import type { DiskPressureStatus } from "../daemon/disk-pressure-guard.js";
-import type { ServerMessage } from "../daemon/message-protocol.js";
 import { setConfig } from "./helpers/set-config.js";
 
 // Short turn-boundary commit wait and no second-pass retitling keep the loop
@@ -134,7 +134,16 @@ function makeCtx(overrides: Partial<Context> = {}): Conversation {
     getQueueDepth: () => 0,
     hasQueuedMessages: () => false,
     canHandoffAtCheckpoint: () => false,
-    drainQueue: async () => {},
+    drainQueue: async (_reason?: string) => {},
+    // Forwards to drainQueue so tests that spy the drain observe the agent
+    // loop's post-turn kick through the guarded entry point.
+    kickDrainQueue(
+      this: { drainQueue: (reason?: string) => unknown },
+      reason: string = "loop_complete",
+      _origin?: string,
+    ) {
+      return this.drainQueue(reason);
+    },
     getTurnInterfaceContext: () => null,
     getTurnChannelContext: () => null,
     buildCurrentSystemPrompt: () => "system prompt",
@@ -157,7 +166,7 @@ describe("runAgentLoopImpl fatal-failure cleanup (ATL-1009)", () => {
   });
 
   test("clears the processing flag before the turn-boundary commit, so a commit that throws cannot latch it", async () => {
-    const events: ServerMessage[] = [];
+    const events: AssistantEvent[] = [];
     // The turn-boundary commit throwing inside the `finally` is precisely the
     // ATL-1007 failure mode. Because the flag is cleared first, the throw can
     // no longer leave `processing_started_at` latched.
@@ -188,7 +197,7 @@ describe("runAgentLoopImpl fatal-failure cleanup (ATL-1009)", () => {
   });
 
   test("clears the processing flag on a clean fatal failure (commit succeeds)", async () => {
-    const events: ServerMessage[] = [];
+    const events: AssistantEvent[] = [];
     const drainQueue = mock(async (_reason: unknown) => {});
     const commitTurnChanges = mock(async () => {});
     const ctx = makeCtx({

package/src/__tests__/conversation-agent-loop-handlers-max-tokens.test.ts

@@ -1,16 +1,16 @@
 import { describe, expect, test } from "bun:test";
 
 import type { AgentEvent } from "../agent/loop.js";
+import type { AssistantEvent } from "../api/index.js";
 import {
   createEventHandlerState,
   type EventHandlerDeps,
   handleMaxTokensReached,
 } from "../daemon/conversation-agent-loop-handlers.js";
-import type { ServerMessage } from "../daemon/message-protocol.js";
 
 describe("max tokens reached handler", () => {
   test("emits and stores an inline continuation card", () => {
-    const sent: ServerMessage[] = [];
+    const sent: AssistantEvent[] = [];
     const ctx = {
       conversationId: "conv-1",
       surfaceState: new Map(),
@@ -18,7 +18,7 @@ describe("max tokens reached handler", () => {
     };
     const deps = {
       ctx,
-      onEvent: (msg: ServerMessage) => sent.push(msg),
+      onEvent: (msg: AssistantEvent) => sent.push(msg),
       reqId: "req-1",
       isFirstMessage: false,
       shouldGenerateTitle: false,
@@ -40,7 +40,7 @@ describe("max tokens reached handler", () => {
 
     const show = sent.find((msg) => msg.type === "ui_surface_show");
     expect(show).toBeDefined();
-    if (!show || show.type !== "ui_surface_show") return;
+    if (!show || show.type !== "ui_surface_show") {return;}
 
     expect(show.surfaceType).toBe("card");
     expect((show.data as { title?: unknown }).title).toBe(

package/src/__tests__/conversation-agent-loop-inference-profile.test.ts

@@ -456,7 +456,16 @@ function makeCtx(
     getQueueDepth: () => 0,
     hasQueuedMessages: () => false,
     canHandoffAtCheckpoint: () => false,
-    drainQueue: () => {},
+    drainQueue: (_reason?: string) => {},
+    // Forwards to drainQueue so tests that spy the drain observe the agent
+    // loop's post-turn kick through the guarded entry point.
+    kickDrainQueue(
+      this: { drainQueue: (reason?: string) => unknown },
+      reason: string = "loop_complete",
+      _origin?: string,
+    ) {
+      return this.drainQueue(reason);
+    },
     getTurnInterfaceContext: () => null,
     getTurnChannelContext: () => ({
       userMessageChannel: "vellum" as const,

package/src/__tests__/conversation-agent-loop-overflow.test.ts

@@ -15,7 +15,7 @@ import { createRequire } from "node:module";
 import { afterAll, beforeEach, describe, expect, mock, test } from "bun:test";
 
 import type { LoopToolExecutor } from "../agent/loop.js";
-import type { ServerMessage } from "../daemon/message-protocol.js";
+import type { AssistantEvent } from "../api/index.js";
 import { resetPluginRegistryAndRegisterDefaults } from "../plugins/defaults/index.js";
 import type { Message, Provider, ToolDefinition } from "../providers/types.js";
 import { ContextOverflowError } from "../providers/types.js";
@@ -595,7 +595,16 @@ function makeCtx(
     getQueueDepth: () => 0,
     hasQueuedMessages: () => false,
     canHandoffAtCheckpoint: () => false,
-    drainQueue: () => {},
+    drainQueue: (_reason?: string) => {},
+    // Forwards to drainQueue so tests that spy the drain observe the agent
+    // loop's post-turn kick through the guarded entry point.
+    kickDrainQueue(
+      this: { drainQueue: (reason?: string) => unknown },
+      reason: string = "loop_complete",
+      _origin?: string,
+    ) {
+      return this.drainQueue(reason);
+    },
     getTurnInterfaceContext: () => null,
     getTurnChannelContext: () => ({
       userMessageChannel: "vellum" as const,
@@ -763,7 +772,7 @@ describe("session-agent-loop overflow recovery (JARVIS-110)", () => {
   test.todo(
     "context too large after progress triggers compaction retry instead of immediate failure",
     async () => {
-      const events: ServerMessage[] = [];
+      const events: AssistantEvent[] = [];
       let reducerCalled = false;
 
       mockReducerStepFn = (msgs: Message[]) => {
@@ -848,7 +857,7 @@ describe("session-agent-loop overflow recovery (JARVIS-110)", () => {
   // loop calibrates the estimator from the rejection and drives the reduction
   // ladder on the next gate pass, recovering before the rerun.
   test("overflow recovery compacts below limit even when estimation underestimates", async () => {
-    const events: ServerMessage[] = [];
+    const events: AssistantEvent[] = [];
     let reducerCalled = false;
 
     // GIVEN the estimator reports 185k and the context manager's compaction
@@ -929,7 +938,7 @@ describe("session-agent-loop overflow recovery (JARVIS-110)", () => {
   test.todo(
     "forced compaction targets a lower budget when estimation has been inaccurate",
     async () => {
-      const events: ServerMessage[] = [];
+      const events: AssistantEvent[] = [];
       let capturedTargetTokens: number | undefined;
 
       // Estimator says 185k (below 190k budget = 200k * 0.95)
@@ -1017,7 +1026,7 @@ describe("session-agent-loop overflow recovery (JARVIS-110)", () => {
   test.todo(
     "overflow recovery succeeds for 75+ message conversation with many tool calls",
     async () => {
-      const events: ServerMessage[] = [];
+      const events: AssistantEvent[] = [];
       const longHistory = buildLongConversation(75);
       let reducerCalled = false;
 
@@ -1099,7 +1108,7 @@ describe("session-agent-loop overflow recovery (JARVIS-110)", () => {
   test.todo(
     "exhausted reducer tiers with progress still attempts emergency compaction",
     async () => {
-      const events: ServerMessage[] = [];
+      const events: AssistantEvent[] = [];
       let emergencyCompactCalled = false;
 
       // Start with reducer already exhausted
@@ -1206,7 +1215,7 @@ describe("session-agent-loop overflow recovery (JARVIS-110)", () => {
   test.todo(
     "onCheckpoint yields when token estimate exceeds mid-loop budget threshold",
     async () => {
-      const events: ServerMessage[] = [];
+      const events: AssistantEvent[] = [];
       let compactionCalled = false;
 
       // estimatePromptTokens is called:
@@ -1307,7 +1316,7 @@ describe("session-agent-loop overflow recovery (JARVIS-110)", () => {
   test.todo(
     "mid-loop budget check prevents context_too_large when tools produce large results",
     async () => {
-      const events: ServerMessage[] = [];
+      const events: AssistantEvent[] = [];
       let compactionCalled = false;
 
       // Budget = 200_000 * 0.95 = 190_000
@@ -1430,7 +1439,7 @@ describe("session-agent-loop overflow recovery (JARVIS-110)", () => {
   test("ladder escalation ends the turn with context_too_large when exhausted", async () => {
     // GIVEN an estimate below the mid-loop threshold, so only the provider's
     // rejection — not the proactive gate — drives recovery
-    const events: ServerMessage[] = [];
+    const events: AssistantEvent[] = [];
     mockEstimateTokens = 100_000;
 
     // AND a ladder that reduces on the first rung and reports exhaustion on
@@ -1497,7 +1506,7 @@ describe("session-agent-loop overflow recovery (JARVIS-110)", () => {
    */
   test("single-rung recovery succeeds and the turn continues in place", async () => {
     // GIVEN an estimate below the mid-loop threshold
-    const events: ServerMessage[] = [];
+    const events: AssistantEvent[] = [];
     mockEstimateTokens = 100_000;
 
     // AND a ladder rung that reduces without reporting exhaustion
@@ -1554,7 +1563,7 @@ describe("session-agent-loop overflow recovery (JARVIS-110)", () => {
    */
   test("multi-rung escalation recovers when a later rung fits", async () => {
     // GIVEN an estimate below the mid-loop threshold
-    const events: ServerMessage[] = [];
+    const events: AssistantEvent[] = [];
     mockEstimateTokens = 100_000;
 
     // AND a ladder that reduces further on each successive rung
@@ -1619,7 +1628,7 @@ describe("session-agent-loop overflow recovery (JARVIS-110)", () => {
    */
   test("budget_yield_unrecovered: classified error emitted, persisted, and stamped", async () => {
     // GIVEN an estimate below the mid-loop threshold
-    const events: ServerMessage[] = [];
+    const events: AssistantEvent[] = [];
     mockEstimateTokens = 100_000;
 
     // AND a ladder whose terminal rung applies `auto_compress_latest_turn` and